.amdb

The Open Standard for AI Context Memory

📐 The Missing Pillar

Is your AI guessing?

Most AI tools (RAG, Vector Embeddings) rely on "fuzzy" matching—finding text that looks similar. This is great for natural language, but terrible for code. Code is precise; dependencies are DAGs, not probability distributions.

.amdb (Agent Memory Database) introduces a Deterministic Context Layer:

• Traditional RAG: "Find code that looks like 'authentication'" → Returns random specialized auth logic mixed with comments.
• .amdb Protocol: "Find the function calling User::login and all its implementations" → Returns the exact Call Graph and Symbol references.

We bridge the gap between Fuzzy Search and Code Structure, giving your AI the "God-mode" accuracy it lacks.

🏗️ Architecture

.amdb runs as a high-performance local daemon, turning your codebase into a queryable knowledge graph.

graph LR
    FS[File System] -->|Notify| Watcher
    Watcher -->|Diff| Parser[Tree-sitter Parser]
    Parser -->|AST| Engine
    Engine -->|Call Graph| DB[(.amdb/store.db)]
    DB -->|Query| MCPServer[MCP Server]
    MCPServer -->|JSON-RPC| Client[Cursor / Claude / Copilot]

1. File Detection: notify watches for real-time changes.
2. Parsing: tree-sitter extracts symbols, ASTs, and imports (Rust, Python, JS/TS supported).
3. DB Storage: Relationships are stored in a local SQLite graph (.amdb/store.db).
4. MCP Server: Exposes data via the Model Context Protocol to any agent.

⚡ Usage Workflow

1. Initialization

Turn any folder into a context-aware project:

amdb init

2. Start the Daemon

Launch the background server to keep context in sync and serve the MCP API:

amdb daemon start

Note: The server binds to 0.0.0.0:3000 by default.

3. Check Status

Verify that your codebase is indexed and the protocol is active:

amdb status

💬 How to use with AI (Crucial)

Since standard AI models rely on embeddings, you must explicitly instruct them to use .amdb for structural accuracy.

📋 Session Start Prompt

Copy and paste this prompt at the beginning of your Cursor/Claude session:

"I am using .amdb in this project. It maintains a deterministic database of symbols and call graphs at .amdb/store.db.

Rule 1: When I ask about code structure (e.g., 'Who calls this?', 'Where is X defined?'), you MUST prioritize the .amdb data over your internal embedding search. Rule 2: If connected via MCP, use the get_context tool first. Rule 3: Trust the .amdb relationships table as the source of truth for dependencies."

🔌 Integration

Cursor

Cursor can natively talk to local servers. Add .amdb as a context source (feature coming soon) or use our generated .cursorrules to guide the model.

Claude Desktop (MCP)

Add the following to your claude_desktop_config.json to enable .amdb as a tool:

{
  "mcpServers": {
    "amdb": {
      "command": "amdb",
      "args": ["mcp", "start"]
    }
  }
}

Now you can ask Claude: "What is the relationship between User and Session structs in this project?"

🗺️ Roadmap

• Phase 1: Local Context Map (Complete)
  • Real-time file watching, Tree-sitter parsing, SQLite Graph storage.
• Phase 2: Semantic Vector Sync (In Progress)
  • Hybrid search combining Deterministic Graph + Semantic Embeddings.

📄 License

This project is licensed under the MIT License.